Towards Automated Eye Movement Characterization for Stroke Patients Using Synthetic Video Data and Machine Learning

Stroke is a critical medical emergency that can cause permanent disability or death. Rapid identification of stroke, especially in prehospital settings, is crucial for timely treatment. Video analysis and machine learning (ML) could facilitate the prehospital assessment of stroke, but a lack of video data from stroke patients remain a barrier to developing effective models. This study explores the use of synthetic data to develop ML models, generating 73 videos mimicking characteristic eye movements of stroke patients through 3D modeling and animation. Four ML models were developed. Long short-term memory (LSTM) and gated recurrent units (GRU) achieved the best performance (over 84% in accuracy, precision, sensitivity, specificity and F1-Score). These findings highlight the promise of synthetic data for developing ML models for healthcare applications and the potential of ML-driven video analysis in the automated assessment of stroke-related eye movements, supporting advancements in prehospital stroke care.